Research On Player Tracking Algorithm In Soccer Match Video

Object tracking is of great significance in soccer videos,which is basis of high-level semantic tasks such as video summary generation,motion analysis,making game strategy and soccer event detection.There are many excellent tracking algorithms have emerged in recent years including correlation filter and deep learning.However,none of them can achieve high accuracy in soccer video.A soccer player tracking algorithm which combines correlation filtering and siamese network has been proposed and designed in this paper.In aspect of feature extraction and processing,the method of combining handicraft features with depth features is adopted in order to enhance expressive ability of features in far-view shot.Discrete features are transformed into continuous features in width and high dimensions by using cubic spline interpolation.Feature depth is greatly reduced by using factorization method to improve tracking speed.As for reducing drifting cases,tracking results are corrected by predicted velocity,which is calculated according to the inertia law of the moving object.In aspect of online tracking,correlation filtering operation is decomposed and added to siamese network as a filter layer.Back-propagation formula of filter layer is deduced for end-to-end training.Second-order Newton iteration method is used to achieve accurate location in sub-pixel level.The method of dynamic learning rate and the strategy of interval frames are adopted for model updating in order to increase adaptability of players' appearance change.The tracker has been tested on soccer data set and compared with other representative trackers.Experimental results show that our tracker has excellent performance in both accuracy and robustness.However,tracking drift is still occurred in occlusion scenes.The next step of research can start with extracting more accurate features and tracking in a way of target division to have a better performance when occlusion.